Selection bias and generalizability

The workplaces included in the TUFS-project, and thereby the study sample, are a result of a strategic sampling procedure. The employees are in no way representative of the working population in Sweden, and this should be kept in mind when drawing conclusions from the findings.

Since the data in the TUFS-project come from six different workplaces it is possible that individuals within one workplace are more similar than individuals from different workplaces. It is also possible that some of the work-related factors studied, have a contextual effect on sick leave on work-unit, workplace or sector-level. Unfortunately, no data is available on work-unit level in TUFS and the number of workplaces and sectors are too few to make multilevel modelling useful (144). However, adjusting for workplace in a regular regression model, as was done in the final model in study I, may imply that the standard errors are underestimated, causing too narrow confidence intervals (144).

The non-participation proportion in Study I, the cohort study, was 53%. Non-participation differed between the occupational sectors, being lowest at the

manufacturing industry. The management and occupational health service staff at this workplace had a strong commitment to work-environment and sick-leave related questions, which may have affected the participation.

For 80% of the non-participants (four of the six participating workplaces), information was available regarding sex and age. The two workplaces where no information on the age and sex of the non-participants was available were both health-care facilities. Of the non-participants (with information available), 56% were women. The age span was 19 to 71 years with a median age of 42 years, which can be compared to 56% women and a median age of 45 years among participants. The sick-leave incidence among the non-participants was 4.30 spells/1 000 person-days, to be compared to 2.85 spells/1 000 person-days among participants.

The non-participation implies a selection of individuals into the study cohort. The higher sick-leave incidence among the non-participants is expected, and may be related to both at health-related selection into the study and an effect of individuals with high levels of sick leave, perhaps to some extent unjustified sick leave, not wanting be contacted during sick leave. To invalidate the results in Study I, participation would have to be differential due to exposure, and high levels instead of low levels of adjustment latitude would have to be related to sick leave among the non-participants.

In total, 432 individuals contributed to the 546 sick-leave spells analysed in Studies II-IV, the case-crossover studies. If this implies a dependency between spells, which may slightly underestimate the variance and possibly affect the risk estimates (145).

However, in none of the studies are the exposed cases confined to spells from a small selected group of participants. There was a median of 79 days between first and second interviews and 88 days between second and third interviews. When restricting the analyses to only first-time interviews, the results did not change markedly.

In 111 of the 679 performed interviews, only a short interview, with no exposure information was conducted. In 198 spells the respondents declined or could not be reached for an interview. Sixty-nine percent of the individuals who participated in a short interview, and 52% of those individuals who declined or could not be reached, contributed with at least one full-length interview during the follow-up for another sick-leave spell.

The non-participation on cohort and interview level mainly affects the generalizability of the results in case-crossover studies. Especially for lack of adjustment latitude, where a surprising decreased risk of sick leave when exposed to lack of adjustment latitude was reported, one cannot assume that the association with sick leave is likely to be similar among non-participants.

The internal non-response varied between the different trigger exposures in the interview. Generally non-response was more common for the trigger exposures where the case period was defined as the first sick-leave day: Non-response was 17-30% for lack of adjustment latitude, 8% for unpleasant work tasks, 23% for a very pressured work situation, and 12-24% for a lower workload than usual. This may imply that the respondents had a hard time estimating exposure on the first sick-leave day, since they were not at work that day. However, the internal non-response also differed depending on the type of control information used, being generally higher when using a weekday- controlled matched pair control period and when using a two-month usual frequency than for other control periods. That missing data is more common in the weekday-controlled analyses are not surprising, since the data was only collected for two weeks prior to sick leave and participants with irregular work schedules may not have worked during a matching weekday during this period. The higher degree of missing data for two-month usual frequency control periods suggest that estimating usual frequency of exposure during such a long period may be difficult for many respondents. The employees who are capable of both estimating the usual frequency over two months and estimating their expected exposure on the first sick-leave day may be a quite selected group for which the general work situation is very similar from day to day.

However, the results from analyses using these data are similar to those from alternative analyses, using different control periods.

Selection bias may arise if case selection is related to trigger exposure in the case period. This could be due to non-participation or flaws in the process of sick-leave reporting. Exposure to problems in workplace relationships, a stressful work situation, bullying or harassment may have been a reason for not participating in the study cohort or for declining to be interviewed during a specific sick-leave spell. Sick-leave spells may have been unreported to the project to evade drawing attention to individuals exposed to psychosocial events. Both these situations may imply that exposed cases are missed, leading to underestimated effect estimates.

Within-individual control period selection bias can arise if an individual experiences a steady up-ward (or down-ward) trend in exposure. Under such circumstances the usual frequency of exposure over a longer period does not reflect the probability of exposure at the time of onset, and the effect estimates will be biased towards infinity (or zero if the trend is downward). In study IV, the two-month usual frequency control periods may have been biased if an decreasing trend in the workload existed over the two-month period, however the effect estimates does not differ notably from those resulting from analyses including other types of control information.

7 CONCLUSION

The overall aim of this thesis was to identify and estimate the effect of factors at work which influence ill individuals to take sick leave. The results indicate that problems in the relationships with superiors or colleagues, and experiencing a very stressful work situation or a lower workload than usual may trigger ill individuals to take sick leave.

The results concerning the effect of lack of adjustment latitude is more ambiguous, suggesting that the measures of adjustment latitude used could capture more than one latent aspect of the work environment.

The broad message of this thesis is that the sick-leave process is not just a matter of identifying risk factors of illness and reduced health. The decision to take sick leave happens in a multifactorial web of circumstances, where health and illness, and their effect on work ability, are not the only important factors. Although illness can be assumed to be strong trigger of sick leave, something else may be the straw that breaks the camel’s back. This thesis sheds some light on what such straws may be.

The results from study II and study IV may imply that sick leave in some instances is used as coping mechanism in very responsible way, that ill individuals tend to plan their sick leave to the days when it does least harm to the organisation. The results also suggest that sick leave may be used by the employer to allow for recuperation when considered possible from a production or organizational point of view, rather than to make efforts to adjust work to be able to make use of employees with temporarily reduced work ability. The overall results suggest that events and situations at work may lower the threshold of work ability reduction, at which an individual feel the need to take sick leave.

Sick leave is not like most health outcomes, where the ultimate goal can be minimizing the outcome incidence. We want to be able to take sick leave, and we want individuals who are too ill to work to report sick. However, interventions to improve workplace relationships and to optimize the workload to the individual employee’s capabilities may have both long-term health-improving effects and, as suggested in this thesis, trigger effects on the decision-making situation of the ill employee.